AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Phrase-Grounding-Aware Supervised Fine-Tuning for Chart Recognition via Side-Masked Attention

Koichiro Ito

CVPR 2026 · Large language models › Instruction and fine-tuning · approach: Efficient LLM fine-tuning systems

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add SMAM phrase-grounding loss to SFTQwen2.5-VL-7B · ChartQA Human splitrelaxed accuracy: 76.2 -> 77.1Improved
add SMAM phrase-grounding loss to SFTLLaVA-7B · ChartQA Human splitrelaxed accuracy: 36.5 -> 38.6Improved
add SMAM phrase-grounding loss to SFTChartInstruct-7B · ChartQA Human splitrelaxed accuracy: 34.5 -> 37.0Improved
add SMAM phrase-grounding loss to SFTQwen2.5-VL-7B · ChartQA Augmented splitrelaxed accuracy: 94.9 -> 95.8Improved
add SMAM phrase-grounding loss to SFTLLaVA-7B · Chart-to-table averageF1: 72.4 -> 74.1Improved
add SMAM phrase-grounding loss to SFT with CoT dataQwen2.5-VL-7B · QA-CoT Human splitrelaxed accuracy: 82.5 -> 83.8Improved
use text localization (bounding-box output) instead of SMAMLLaVA-7B · Chart-to-table averageF1: 72.4 -> 36.0 (vs SFT)Hurt
use text localization with CoT dataQwen2.5-VL-3B · QA-CoT Human splitrelaxed accuracy: 78.9 -> 75.2Hurt
use segmentation-based GLaMM-style groundingLLaVA-7B · Chart-to-table averageF1: 72.4 -> 18.9Hurt
use text localization on ChartQAQwen2.5-VL-7B · ChartQA Human splitrelaxed accuracy: 76.2 -> 76.4No clear effect
add grounding loss L_G alongside L_CE+ (gamma 0.5, alpha 0.1) (ablation)SMAM on Qwen2.5-VL-7B · ChartQA Human splitrelaxed accuracy: 76.8 -> 77.1Improved
add L_CE+ regularization to L_G (alpha 0.1) (ablation)SMAM on LLaVA-7B · ChartQA Human splitrelaxed accuracy: 37.0 -> 38.6Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy